Sequence-based visual place recognition: a scale-space approach for boundary detection

In the field of visual Place Recognition (vPR), sequence-based techniques have received close attention since they combine visual information from multiple measurements to enhance the results. This paper is concerned with the task of identifying sequence boundaries, corresponding to physical scene limits of the robot's trajectory, that can potentially be re-encountered during an autonomous mission. In contrast to other vPR techniques that select a predefined length for all the image sequences, our approach focuses on a dynamic segmentation and allows for the visual information to be consistently grouped between different visits of the same area. To achieve this, we compute similarity measurements between consecutively acquired frames to incrementally formulate a similarity signal. Then, local extrema are detected in the Scale-Space domain regardless the velocity that a camera travels and perceives the world. Accounting for any detection inconsistencies, we explore asynchronous sequence-based techniques and a novel weighted temporal consistency scheme that strengthens the performance. Our dynamically computed sequence segmentation is tested on two different vPR methods offering an improvement in the systems' accuracy.